Mircea Giurgiu * Results on Automatic Speech Recognition in Romanian

A vector quantizer is a system for mapping a sequence of continuous or discrete vectors into a digital sequence suitable for communication over or storage in a digital channel. The goal of such a system is data compression. In the past few years several algorithms have been designed for VQ and their performance has been studied for speech wave. A k-dimensional memoryless VQ consists in two mappings: (a) an encoder, which assigns to each input vector X=[x(0), x(1),...,x(k-1)] a channel symbol in the symbol channel set: {Y(i); i=1,2,...,M}, and (b) a decoder , which assigns to each channel symbol "v" a value in the reproduction alphabet A [1,6,7]. To set up a VQ system means to realize a partition in the set of all vectors X in classes C_i. In each of them it distinguishes a particular vector Y_i called a prototype or "centroid". Each vector X in C_i will be represented by his prototype Y_i. All set of Y_i realize a codebook (Fig. 5). Substituting X with Y realizes a distortion d(X, Y).

Fig. 5. -Vector Quantization principle.

The most important problem in VQ is the design of C from a set T of training vectors. Two sets of experiments have been designed by using Linde-Buzo-Gray (LBG) algorithm [7]

SET1: 6613 vectors collected from 1 speaker uttering 10 times the 10 digits (Table 1)
SET2: 6613 vectors collected from 3 speakers uttering 3 times the 10 digits (Table 2)

and we generated vector quantizers of size M =2, 4, 8, 16, 32, 64, and 128. During the course of running the algorithms (Fig. 6-9), several performance criteria were monitored including: (a) average distortion (SNR versus number of iterations); (b) cluster separation of the resulting coding entries; (c) cluster cardinality (number of tokens in each cluster); (d) cluster distortion; (e) SNR in the encoding process (after VQ training) [3].

Table1. VQ performances for SET1

VQ size M	SNR Training [dB]	Cluster Cardinality	SNR Coding [dB]
32	24.9	209.35	23.35
64	26.4	104.07	26.88
128	27.8	51.82	28.25

Table 2. VQ performances for SET2

VQ size M	SNR Training [dB]	Cluster Cardinality	SNR Coding [dB]
32	24.9	225.06	25.27
64	26.2	111.92	26.56
128	27.5	55.90	27.91

Table 3. SNR versus the VQ size (M) for the word /unu/ ("one") - SET1

VQ size M	SNR Coding [dB]
32	24.33
64	25.40
128	26.70

The conclusion we found for VQ is that it is a powerful tool in speech processing: it produces a high compression rate and reduces the amount of computation in the ASR process. The technique has been applied for isolated word recognition using DTW and also with HMM. For values of M greater than 32 the distortion error does not decrease very much, so we select M=64 in our approaches. The larger cluster has 362 tokens, whereas the smaller cluster has 62; hence, a spread of over 7 to 1 in cluster occupancy is obtained. The average cluster cardinality is 225 (for the case M=32) and the histogram of cluster cardinality indicates that the vast majority of clusters have fewer than the average number of tokens. A spread of more than 6 to 1 was also observed for cluster distortion [3].

180